ICRA 2026

IndustryShapes dataset

An RGB-D Benchmark dataset for 6D object pose estimation of industrial assembly components and tools

Panagiotis Sapoutzoglou Orestis Vaggelis Athina Zacharia Evangelos Sartinas Maria Pateraki
National Technical University of Athens, Greece
ArXiv Paper Code Dataset (Hugging Face)

Industrial textured 3D Models

Obj 1 Obj 2 Obj 3 Obj 4 Obj 5

Data from Real Industrial Environment

Multiple Instances
Multiple Instances
Occlusion and Clutter
Occlusion & Clutter

RGB-D Static Onboarding sequences

Rotating 3D RGB
RGB Sequence
Rotating 3D Depth
Depth Sequence

Overview

IndustryShapes is a new benchmark dataset tailored for 6D object pose estimation in industrial settings. Targeting the challenges of textureless objects, reflective surfaces, and complex assembly tools, this dataset provides high-quality RGB-D data with precise annotations to advance the state of the art in robotic manipulation.

5 Industrial Objects
22.4k Total Annotations
>6.6k RGB-D Images
23 Total Scenes

Structure & Composition

Classic Set

The Classic Set supports instance-level pose estimation with 21 scenes (13 train, 8 test).

  • Real Industrial Scenes: Assembly settings with varying complexity (single to cluttered/occluded).
  • Lab Capture: Systematic sampling using a turn-table setup.
  • Synthetic Data: OpenGL-rendered images with photorealistic textures.

Captured with Intel RealSense D455 (640x480).

Extended Set

The Extended Set supports the benchmarking of novel pose estimation methods (model-based & model-free).

  • Onboarding Sequences: 10 RGB-D static onboarding sequences (2 per object).
  • Challenging Office scenes: Three scenes captured in an office environment with unconstrained lighting, distractors, occlusions and diverse viewpoints featuring all objects.

Captured with Intel RealSense D405 for close-range depth. Adds ~10.3k annotated instances and dense camera viewpoints.

Distributions & Analysis

The dataset features a balanced distribution of object-to-camera distances (mostly 400-800mm) and dense camera orientation coverage.

Orientation Distribution Camera Orientation Distribution
Distance Distribution Object-to-Camera Distance Distribution

Citation

@inproceedings{sapoutzoglou2026industryshapes,
  title={IndustryShapes: An RGB-D Benchmark dataset for 6D object pose estimation of industrial assembly components and tools},
  author={Sapoutzoglou, Panagiotis and Vaggelis, Orestis and Zacharia, Athina and Sartinas, Evangelos and Pateraki, Maria},
  booktitle={ArXiv}, 
  year={2026}
}